Novel machine learning model can predict material failure before it happens

(Press-News.org) A team of Lehigh University researchers has successfully predicted abnormal grain growth in simulated polycrystalline materials for the first time—a development that could lead to the creation of stronger, more reliable materials for high-stress environments, such as combustion engines. A paper describing their novel machine learning method was recently published in Nature Computational Materials. 

“Using simulations, we were not only able to predict abnormal grain growth, but we were able to predict it far in advance of when that growth happens,” says Brian Y. Chen, an associate professor of computer science and engineering in Lehigh’s P.C. Rossin College of Engineering and Applied Science and a co-author of the study. “In 86 percent of the cases we observed, we were able to predict within the first 20 percent of the lifetime of that material whether a particular grain will become abnormal or not.”

When metals and ceramics are exposed to continuous heat—like the temperatures generated by rocket or airplane engines, for example—they can fail. Such materials are made of crystals, or grains, and when they’re heated, atoms can move, causing the crystals to grow or shrink. When a few grains grow abnormally large relative to their neighbors, the resulting change can alter the material’s properties. A material that previously had some flexibility, for instance, may become brittle.

“As we design materials, we’d like to be able to design them intentionally to avoid abnormal grain growth,” says Chen.

A smarter way to identify stable materials

To date, however, predicting abnormal grain growth has been a needle-in-a-haystack problem. There are countless combinations and concentrations that can go into the creation of any given alloy. Each of those metals must then be tested, which is expensive, time-consuming, and often impractical. The computational simulation developed by Chen’s team helps narrow down possibilities by quickly eliminating materials that are likely to develop abnormal grain growth.

“Our results are important because if you want to look at that big haystack of different materials, you don’t want to have to simulate each one for too long before you know whether or not abnormal grain growth is going to occur,” he says. “You want to simulate for as little time as possible, and then move on.”

The challenge is that abnormal grain growth is a rare event and, early on, the grains that will become abnormal look just like the others.

Unlocking hidden patterns with AI

To address this, the team developed a deep learning model that combined two techniques to analyze how grains evolve over time and interact with each other: A long short-term memory (LSTM) network modeled how the properties—or features—of the material would be evaluated and a graph-based convolutional network (GCRN) established relationships between the data that could then be used for prediction. 

Initially, the researchers simply hoped to make successful predictions. They didn’t anticipate being able to make predictions so early.

“We thought that the data might be too noisy,” he says. “Maybe the properties we were looking at wouldn’t reveal very much about distant future abnormalities, or maybe the abnormality would only reveal itself just as it was about to happen, when it might be obvious even to the human eye. But we were surprised that we were actually able to make predictions so far in advance.”

Critical to that early detection was using their models to examine the grain’s characteristics over time before the abnormality occurred. 

“A better way to think about grains becoming abnormal is to think about how they evolve in the time before they change,” he says. “So at 10 million time steps before abnormality, for example, they have certain properties that might differ from those they had at 40 million time steps.”

The team aligned each simulation at the point in time where the grain became abnormal, and worked backward examining its evolving properties. By identifying consistent trends in these properties, they were able to accurately predict which grains would become abnormal.

“If you look at the grains in terms of how much time before they transition, you can see shared trends that are useful for prediction,” he says.

In this project, Chen and his team conducted simulations of realistic materials. The next phase is to apply the approach to images of real materials and see if they can still accurately predict the future. The ultimate goal, says Chen, is to identify materials that are highly stable and can maintain their physical properties under a wide range of high-temperature, high-stress conditions. Such materials could allow engines and engine parts to run at higher temperatures for longer before failure.

The team also sees the potential of their novel machine learning method to predict other rare events, both within and beyond the field of materials science, thanks to its ability to identify warning signs in complex systems. For example, it could potentially help predict phase changes in materials, mutations leading to dangerous pathogens, or sudden shifts in atmospheric conditions.

“This work opens up an exciting new possibility for material scientists to ‘look into the future’ to predict the future evolution of material structures in ways that were never possible before,” says Martin Harmer, Lehigh’s Alcoa Foundation Professor of Materials Science and Engineering, Emeritus; director of the Nano/Human Interface Presidential Research Initiative; and co-author of the paper. “It will have a major impact in designing reliable materials for defense, aerospace and commercial applications.”

Research Team and Funding

Computer science and engineering PhD student Houliang Zhou and MS student Benjamin Zalatan co-authored the paper (npj Comput Mater 11, 82 (2025)) along with Chen, Harmer, and Joan Stanescu, visiting scholar in Lehigh’s Nano | Human Interfaces (NHI) Presidential Initiative; Jeffrey M. Rickman, Class of ’61 Professor of Materials Science and Engineering; Lifang He, an associate professor of computer science and engineering; and Lehigh alum Christopher J. Marvel ’12 ’16 PhD, an assistant professor of mechanical engineering at Louisiana State University. 

This work was supported in part by the National Science Foundation, the Army Research Office, the Army Research Laboratory Lightweight High Entropy Alloy Design (LHEAD) Project, and the Lehigh University Presidential NHI Initiative.

Related figures and videos of the simulation (see Supplementary Information)  are available via the Open Access article in npj Computational Materials. 

Related Links

npj Computational Materials: "Learning to predict rare events: the case of abnormal grain growth" Rossin College Faculty Profile: Brian Y. Chen Lehigh NHI Presidential Initiative: Martin Harmer Lehigh NHI Presidential Initiative: Joan Stanescu Rossin College Faculty Profile: Jeffrey M. Rickman Rossin College Faculty Profile: Lifang He Louisiana State University: Christopher Marvel END